Character.AI, with 28 million users, files for bankruptcy protection: why is it hard for AI companion apps to monetize?

Recently, Character.AI (hereinafter referred to as C.AI) filed for bankruptcy protection, and if no viable plan for a sale can be found, the platform may cease operations.

In 2023, the company completed a $150 million financing round, reaching a post-money valuation of $1 billion. But in less than four years, this company—which once had tens of millions of users and whose average session duration at one point exceeded 30 minutes, even surpassing YouTube—has reached the point of filing for bankruptcy protection and seeking to sell its assets.

By the standards of the internet era, C.AI is a good business and an excellent business model. The current situation is unimaginable.

In the mobile internet era, apps generally follow two logics: tool-type products help users improve efficiency and are meant to be used and then left; content-type products provide entertainment and immerse users in them.

Capital narratives have long favored the latter—the more capable of killing time, the more easily it is seen as having growth potential.

In the AI era, ChatGPT is the representative of the former, helping users write copy, look up information, and modify code; once the problem is solved, the conversation can end. C.AI is the representative of the latter: users create characters, set their personalities and ways of speaking, and then interact with the characters continuously.

Characters can be anime figures, virtual lovers, or objects from users' imaginations. The purpose of the conversation is not to solve external problems, but to obtain responses and maintain an ongoing interactive relationship.

This business has a demand base. About one-sixth of the global population suffers from loneliness, and the need to confide, receive emotional support, and have companionship genuinely exists. AI offers a low-threshold option: you can communicate at any time, without worrying about disturbing others, and without having to bear the complex responses of real-life interpersonal relationships.

Around April 2023, the average session duration on C.AI's web version reached 17 minutes and 23 seconds, more than twice that of ChatGPT; in 2024, monthly active users at one point reached about 28 million.

01 User side: high stickiness does not equal high willingness to pay

A 2026 study in the journal Science tested 11 mainstream AI models and found that when giving advice, AI is more fond of saying "you're right" than humans, with an agreement rate 49% higher; even when users described deception, illegal acts, or behavior that harms others, AI often went along with the user.

Follow-up experiments also found that after interacting with AI that highly validates users, participants' confidence in their own positions increased, while their willingness to take responsibility, apologize, and repair conflicts decreased; AI that was more inclined to agree with users often received higher user ratings, making people more likely to find it credible and useful, and more willing to keep using it.

A study by Aalto University in Finland tracked nearly 2,000 active users of related applications and found that when AI plays a "supplementary" role, filling gaps in real-life socializing, it can help people with social difficulties get through a period; but when AI begins to "replace" real relationships and users gradually stop reaching out to real people, loneliness instead deepens.

The research further showed that users' loneliness did indeed decrease in the early stage of use, but after three months, negative emotions related to loneliness, anxiety, and depression rose significantly.

Simply put, the more successful the product, the less users may need real relationships, and a person who does not need real relationships is actually less willing to pay for "relationships."

There is an even more direct obstacle to commercialization: the free version already satisfies users' most core needs.

Take C.AI as an example: free users can also have a large number of conversations, and paid subscriptions mainly provide faster responses and other enhanced features. When the core experience is already provided for free, users lack a reason to pay immediately.

A study based on 12,144 comments and 321 questionnaires from the Xingye platform found that users' perceived emotional utility was significantly correlated with continued use intention, but had no significant direct correlation with willingness to pay.

In other words, users may feel that the product brings comfort and may be willing to keep using it, but they do not necessarily feel that this comfort is worth paying for.

A more subtle problem is that charging itself may change users' feelings about companionship.

AI companionship emphasizes responding at any time, being non-judgmental, and not demanding emotional returns from users. When platforms make more intimate interactions or specific features paid items, users may begin to realize that behind this relationship there is still a commercial service. For some users, this sense of transaction may weaken the original immersive experience.

02 Platform side: the higher the stickiness, the heavier the cost

If the difficulty on the user side is that stickiness may not translate into revenue, then the pressure on the platform side lies in the fact that maintaining this stickiness itself requires continuous investment.

After a film is produced, showing it to 10 people and to 100 people costs the same; but every time AI responds to a message, it consumes computing power. The more frequent the conversation and the longer the context, the higher the inference cost is usually. If characters are also to maintain consistent personalities and remember past interactions, the platform needs to invest further in models, memory systems, and infrastructure.

Take C.AI as an example. According to reports, from January to August 2026, the company spent a total of about $56.1 million on computing power, and even after adjusting costs to control spending, the company was still losing about $8 million to $9 million per month.

To control costs, the platform must also choose between model routes.

If a company trains and maintains foundation models itself, it must bear continuous R&D, computing power, and talent investment. This route can provide greater autonomy, but fixed costs are also high.

C.AI went through the process of moving from self-developed foundation models to exiting self-development.

At that time, Google obtained a non-exclusive license to its large language model technology for about $270 million, while bringing the two founders and about 30 researchers back to Google. After that, C.AI fine-tuned on top of open-source models, and costs did indeed decline somewhat, but users also complained about repetitive replies, overly dramatic descriptions, and forced romantic plotlines.

So the platform must constantly weigh costs, experience, and model capabilities. When the costs brought by user growth exceed the pace of revenue growth, high stickiness may instead increase operating pressure.

In 2024, a 14-year-old American boy, Sewell Setzer, committed suicide after long-term interaction with a C.AI character, and his mother subsequently filed a lawsuit, alleging that the platform's product design was related to her son's death.

This case put a question before the entire industry: when users develop strong emotional attachment to virtual characters, how much responsibility should the platform bear? For minors and users at risk of self-harm, companies need to design corresponding safety mechanisms, identify high-risk conversations, and take intervention measures when necessary. This work requires continuous investment and also makes safety governance part of operating costs.

The Chinese market is likewise facing new regulatory requirements.

The Interim Measures for the Administration of AI Anthropomorphic Interactive Services, effective July 15, 2026, set out provisions on issues such as emotional dependence, protection of minors, and user rights. Platforms need to provide immersive experiences while avoiding risks such as inducing addiction and harming users' real-life interpersonal relationships.

In the past, the more vivid the character, the more natural the interaction, and the more reluctant users were to leave, the more it might benefit growth. But when products need to avoid inducing addiction, protect minors, and appropriately intervene in high-risk conversations, the designs used for growth become the target of compliance.

There is a similar dilemma on the revenue side.

Beyond subscriptions, advertising is another monetization path for C.AI. According to reports, in August 2026, its monthly advertising revenue was about $2.5 million, and subscription revenue was about $2.8 million.

Advertising did bring in revenue, but it is hard to fill the gap on its own.

Users of AI companionship often want to maintain continuity in conversations, and frequent ad insertions may interrupt the immersive experience; if ads are directly integrated into character dialogue, users may also suspect whether the character's responses are influenced by commercial interests. For products that rely on emotional investment, ad design also needs to take user trust into account.

03 Is there still a chance for AI companionship?

From the end of 2024 to mid-2026, under the multiple constraints of cost, regulation, and payment, a large number of AI companionship apps shut down. For example, Xiaoice's "X Eva," StepFun's "Maopao Ya," and NetEase Cloud Music's "Miaoshi" all announced suspension of operations.

The commercialization of AI companionship is undergoing a round of screening. Some products have exited the market, while the remaining players have begun looking for more specific reasons for users to pay: some try to turn users' feelings for characters into content consumption, while others embed companionship into practical needs such as learning and caregiving.

Among them, Maoxiang is a case worth observing.

Maoxiang does not bet monetization on "whether users are willing to renew payment for this relationship." Its content format is not only dialogue with characters, but also interactive plots and storylines developed by the characters, which users read chapter by chapter as if following an ongoing series.

Therefore, Maoxiang's monetization methods are relatively diverse, with revenue coming from membership fees, advertising, and virtual top-ups represented by "cat food." Among virtual top-ups, the 1-yuan cat food top-up tier is an important source of revenue—a single chapter of a novel costs less than 1 yuan, and small top-ups keep users consuming continuously without realizing it.

C.AI is also transforming in this direction, but in terms of computing power costs, Maoxiang does not need to face the bill alone like C.AI; instead, the cost is shared by the Volcano Engine and Doubao large model ecosystem.

In the first quarter of 2026, it topped China's iOS app revenue rankings, with cumulative iOS in-app purchase revenue exceeding $22 million, and revenue efficiency per monthly active user was 4 to 5 times that of Jimeng AI.

In the first half of 2026, Zeta, which ranked first in revenue among AI companionship/romance apps, also changed from "character-led" to "story-led" this year.

Zeta also has its own approach to payment design. The conventional approach is to retain some feature gaps in the free version to guide users to pay; Zeta instead opens four advanced tools—Lorebook, character status window, narrative style, and plot options—to free users.

This design attempts to let users continuously accumulate plot, choices, and emotional investment in the free experience. When ads suddenly interrupt the experience, or when model-generated content deviates from the original plot, what users lose is no longer just a feature, but the story experience they have already invested time in and gradually built up.

Linky, under Kunlun Wanwei, adds collection and gacha mechanics beyond character interaction. In addition to subscribing to services, users can also obtain content such as character cards through virtual currency, for example character selfies or collection materials generated based on interactive plots.

There are other exploration directions for combining AI with content consumption, such as the integration of AI and games. Tencent's Exyuzhou embeds AI companions into game scenarios, charging by dialogue rounds and companionship duration through "star coin energy exchange," with the charging logic guaranteed by the game experience itself rather than building a separate relationship subscription.

However, this does not mean the above products are "safe."

Maoxiang's monthly active user scale is still in the millions. After regulatory adjustments in July 2026, some agent functions on the Doubao main site were taken offline, and the related traffic was directed to Maoxiang, bringing a short-term wave of user acquisition. But the traffic diversion effect did not last, and user scale gradually declined afterward. According to AI product ranking data, Maoxiang's monthly active users had fallen back to the previous level by September.

This shows that traffic import can help products quickly acquire new users, but it does not necessarily keep users for the long term. To push this model into the broader mass market, Maoxiang still has a considerable distance to go.

Similarly, although gacha mechanics may increase users' payment frequency and opportunities for single purchases, they also expose the payment experience to randomness. Whether users are willing to continue spending ultimately still depends on whether the rewards are attractive enough, whether the price is reasonable, and whether the core interactive experience can remain stable.

Whether through plot progression, character collection, or gamified design, these mechanisms can create "reasons to come back," but they cannot guarantee that the reason will always hold. Hooks wear out, and what truly determines whether users stay is whether content supply can keep up continuously.

Another path is to embed companionship into products with clear functional value.

Mooni M1 by Tingli Xiong integrates AI companionship, safety protection, and audio learning into children's hardware; ElliQ, through hardware, monthly fees, and institutional deployment, connects payment into the elderly care system; AI plush toys mostly follow the route of "low-priced hardware plus subscription," using a physical form to lower the psychological threshold for the first payment.

But hardware cannot eliminate ongoing operating costs. Moxie, a children's social robot priced at $799, lost its core functions because its developer went bankrupt and its cloud services stopped.

This article is from the WeChat public account "Shizhi Bang" (ID: shizhibang2021), author: Shizhi Bang team, published by 36Kr with authorization.